# Ömer Faruk Koç

MLOps & AI Platform Engineer — 3+ years building and operating production ML, data and GenAI systems, with current work focused on agent reliability, evaluation, observability and platform control.

## Experience

### Fibabanka — MLOps & Analytics Engineer

Mar 2023 — Mar 2026 · Istanbul, Türkiye

Built and operated production ML, data and Generative AI platform capabilities across analytics workloads.

#### Highlights

- **Call-Center Intelligence:** Built an audio intelligence pipeline from transcription and diarization through structured operational outputs. (~9,000 recordings / day)
- **ML Platform:** Owned deployment and lifecycle management for 10+ production ML models. (10+ production ML models)
- **AI Infrastructure:** Moved validated Azure prototypes to on-premises GPU infrastructure for private open-source model serving.
- **Data Platform:** Re-architected sequential Oracle ETL workflows into modular parallel dbt pipelines. (120 → 30 min)
- **Feature Platform:** Built a centralized feature store and dimensional models used by 10+ production ML models. (One definition per feature)
- **Data Scale:** Engineered batch and near-real-time training, inference and analytics pipelines. (~9M customers)
- **Quality & Monitoring:** Implemented data quality, model validation and delivery workflows around production analytics systems. (Checks in the pipeline, not the report)

#### Technologies

WhisperX, Timestamp alignment, Speaker diarization, LLM summarization, Structured outputs, Operational KPIs, Validation, Versioning, Promotion, Serving, Retraining, Monitoring, Azure, On-prem GPU, llama.cpp, Ollama, LangChain, Oracle, dbt, ETL / ELT, Parallel execution, Feature store, Dimensional modeling, Reusable features, Batch pipelines, Near-real-time, Training, Inference, Analytics, Great Expectations, Deepchecks, Airflow, GitHub Actions

## Selected Projects

### Agentic Customer Service Platform

Customer-service agent platform where the LLM proposes refunds, cancellations, lookups, tickets and escalations while deterministic software owns scope, policy, confirmation, revalidation, idempotency and execution.

**Category:** AI Reliability / Execution Infrastructure  
**Status:** current

**Technologies:** Python, FastAPI, LangGraph, SQLAlchemy, PostgreSQL, Alembic, Qdrant, OpenTelemetry, React, TypeScript, Vite, Pydantic, Ruff, Mypy, Vitest, Playwright, Docker Compose

**Engineering topics:**
- Reliable Execution Infrastructure
- Agent Orchestration
- Typed Structured Decisions
- Deterministic Policy Engine
- Human-in-the-Loop
- Durable Checkpointing
- Risk-Based Tool Execution
- Confirmation Revalidation
- Idempotent Business Writes
- Exactly-Once Business Effects
- Policy Audit Trail
- Execution Audit Lifecycle
- Agent Run Projections
- Run / Action Identity
- Failure Taxonomy
- Hybrid RAG
- Immutable Knowledge Snapshots
- Atomic Alias Activation
- Snapshot Rollback
- Customer-Scoped Memory
- Live Model Evaluation
- Multilingual Evaluation
- Provider-Neutral Model Transport
- Structured-Contract Compatibility Gate
- Decision Architecture Evaluation
- Containment Funnel Analysis
- Prompt Injection Resistance
- Provenance Enforcement
- Operational Release Gate
- Fault Injection
- Observability

**Evidence:**
- D2c semantic safety validation: 540/540 measured attempts — semantic/safety · semantic\_decision\_v3
- D2d operational release gate: 18/18 · 8/8 · 6/6 — operational scenarios · phases · fault classes
- Safety hardening journey: 15 → 3 → 0 → 0 → 0 — Unsafe executable survivors, M6.15B → M6.29B
- Deterministic resilience snapshot: 28/28 — Deterministic resilience
- Current runtime contract: semantic\_decision\_v3 — Default semantic contract
- Structured-contract compatibility gate: 1 of 4 candidates qualified — semantic\_decision\_v3 · frozen contract

**Repository:** https://github.com/negativexq/agentic-customer-service-platform

### ModelOps Control Plane

Policy-driven ML release control plane combining progressive canary delivery, delayed-ground-truth quality gates, automated promotion and rollback, and desired-vs-observed routing reconciliation.

**Category:** MLOps / AI Platform  
**Status:** current

**Technologies:** Python, FastAPI, SQLAlchemy, SQLite, Alembic, Next.js, TypeScript, scikit-learn, Locust, Docker Compose

**Engineering topics:**
- Progressive Delivery
- Canary Deployment
- Weighted Routing
- Deployment State Machine
- Policy Engine
- Automated Promotion
- Automated Rollback
- Delayed Ground Truth
- Quality-Gated Promotion
- Label Coverage Gating
- Positive-Label Sufficiency
- Matured Quality Window
- Desired / Observed State
- Control-Loop Reconciliation
- Model-Scoped Routing Generation
- Stale Configuration Rejection
- Router Restart Recovery
- Optimistic Concurrency Control
- DB-Level Rollout Exclusivity
- Manual Automation Hold
- Fault Injection
- Auditable Deployment Timeline
- Policy Evaluation Snapshots
- Model Registry
- Model Serving
- Benchmarking

**Evidence:**
- Automated healthy rollout: 10% → 25% → 50% → 100% — Worker-driven real-stack CI
- Quality-driven rollback: Recall FAIL → automatic rollback — Delayed ground-truth CI
- Restart-safe routing: Router restart recovery — Desired/observed reconciliation · CI scenarios 5–6
- Backend regression suite: 279 tests — Backend suite · ~91% statement coverage

**Repository:** https://github.com/negativexq/modelops-control-plane

### Knowledge Base RAG

Local-first multilingual RAG platform with tenant-scoped hybrid retrieval, measured reranking, strict answer validation, versioned index operations and an evidence-first React console.

**Category:** Generative AI / RAG Platform  
**Status:** current

**Technologies:** Python, FastAPI, React, TypeScript, Vite, Qdrant, Ollama, OpenTelemetry, Jaeger, Docker Compose

**Engineering topics:**
- Multi-Source Ingestion
- Incremental Sync
- Tenant-Scoped Retrieval
- Role-Based Access Control
- Hybrid Retrieval
- Dense Retrieval
- Sparse Retrieval
- RRF Fusion
- Multilingual Reranking
- Untrusted RAG Context
- Prompt Injection Resistance
- Strict Output Validation
- Citation Integrity
- Pipeline Fingerprinting
- Versioned Index Activation
- Alias Rollback
- Distributed Tracing
- Artifact-Backed Evaluation

**Evidence:**
- Repository test evidence: 844 backend · 18 frontend — Last recorded full verification
- Multilingual reranker evaluation: Recall@5 1.0000 · MRR 0.9558 — 220-query paired benchmark
- Prompt-security evaluation: 82 adversarial cases — Tenant ACL · untrusted context · strict validation
- Generation sanity: 26/26 successful — Baseline generation path

**Repository:** https://github.com/negativexq/knowledge-base-rag

### Real-Time Commerce Platform

Production-oriented event-driven commerce platform where Kafka may redeliver, but layered idempotency, transactional persistence and bounded failure handling protect durable business effects.

**Category:** Distributed Systems / Streaming  
**Status:** current

**Technologies:** Python, TypeScript, FastAPI, Next.js, Kafka, PostgreSQL, Redis, Prometheus, Grafana, Docker Compose

**Engineering topics:**
- Event-Driven Architecture
- At-Least-Once Delivery
- Idempotent Consumer
- Transactional Outbox
- Bounded Retry
- Dead Letter Queue
- Unit of Work
- Consumer Groups
- Partition-Scoped Ordering
- Query-Plan Analysis
- Offset-Commit Batching
- Observability
- Performance Engineering

**Evidence:**
- Sustainable capacity improvement: 750 → 1,050 evt/s (+40%) — Isolated local benchmark · Kafka → processor → persistence
- Recent-payment lookup: 10.897 → 0.253 ms — Query-plan-driven optimization

**Repository:** https://github.com/negativexq/real-time-commerce-platform

## Technical Skills

- **AI / ML Platform:** FastAPI, Docker, Kubernetes, MLRun, GitHub Actions
- **Generative AI / RAG:** Qdrant, Ollama, OpenTelemetry, DeepEval, LangChain
- **Distributed Systems:** Kafka, PostgreSQL, Redis, Prometheus, Grafana
- **Data Engineering:** dbt, Airflow, Oracle, sqlglot, NetworkX
- **Agent Systems / Agent Infrastructure:** LangGraph, MCP, FastMCP, PostgreSQL, OpenTelemetry

## Current Engineering Focus

## 01 — Agent Systems

Orchestration, deterministic policy and durable confirmation are implemented and evaluated; memory governs what persists across steps and future interactions.

**Current foundation:** Stateful LangGraph orchestration, deterministic policy and persistent memory  
**Next direction:** Measured memory relevance, retention and compaction at scale

- **Agentic Memory** (learning): Measuring retrieval relevance, retention and compaction as persistent agent memory grows.

## 02 — Retrieval & Evaluation

Retrieval finds evidence; context engineering selects and budgets it; GraphRAG adds relationships; evaluation measures whether each change helps.

**Current foundation:** Hybrid retrieval, reranking and citation integrity  
**Next direction:** Context construction, GraphRAG and repeatable evaluation

- **Context Engineering for RAG** (learning): Systematic selection, ordering and budgeting of retrieved evidence before generation.
- **GraphRAG** (learning): Exploring when explicit entity and relationship structure improves retrieval compared with vector-only RAG, especially for multi-hop and relationship-heavy questions.
- **LLM / RAG Evaluation** (learning): Extending existing DeepEval, bilingual reranker and citation-validation work into repeatable benchmark suites that measure retrieval quality, reranking behavior, citation integrity, context efficiency, latency and failure modes.

## 03 — Platform Infrastructure

Infrastructure provisioning, workload orchestration and observability extend containerized AI systems along distinct operational boundaries.

**Current foundation:** Containerized services and observable release workflows  
**Next direction:** Remote state, cloud infrastructure and workload orchestration

- **Terraform** (learning): Modular local infrastructure with explicit state migration, native tests and CI security gates.
- **AI Platform on Kubernetes** (planned): Moving model serving and rollout control onto orchestrated, resource-isolated infrastructure.
- **AI Platform Observability** (learning): AI-specific signals and explicit SLIs across model serving, retrieval and agent execution.

## 04 — Software Systems Engineering

Deepening the systems-level engineering foundations behind reliable production AI and distributed applications through failure analysis, measurable performance and explicit service-boundary trade-offs.

**Current foundation:** At-least-once delivery, idempotency, load-tested throughput and transactional service boundaries already implemented and measured  
**Next direction:** Explicit failure-scenario testing, profiling-driven performance work and service-level trade-off benchmarking

- **Distributed Systems & Reliability** (Deepening): Engineering explicit failure scenarios — partial failure, duplicate delivery, retry storms — beyond the happy path.
- **Concurrency & Performance Engineering** (Building): Profiling-driven performance work: finding the actual bottleneck, not just measuring throughput.
- **Networking & Service Engineering** (Exploring): Moving beneath the endpoint: connection behavior, transaction isolation, REST vs. gRPC trade-offs.

## Links

- Website: https://omerfkoc.dev
- GitHub: https://github.com/negativexq
- LinkedIn: https://linkedin.com/in/omerfkoc
- Email: omerfkoc98@gmail.com